Flare Flow

Problem Statement

Most of our business logic runs in Cadence on Flow. The frontend is built with TypeScript, HTML, and CSS. We created a Flare connector (in Cadence) that streams per-key LiteLLM usage back on-chain so the contract can govern keys. Flow’s transaction model lets us encode access controls across multiple actions and execute them atomically in a single tx.We also provide a Docker Compose setup around the open-source liteLLM proxy so all model calls go through a controlled endpoint; the proxy runs inside a TEE for added isolation.

Solution

On Flow, we built our Cadence smart contracts to run on-chain billing and access control for AI models. Each model gets a resource-backed vault funded in FLOW; the contract issues OpenAI-compatible API keys with optional expirations and scoped permissions. A Cadence/Flare connector streams per-key LiteLLM usage back on-chain via events, enabling real-time metering and threshold enforcement. Flow’s capability system and atomic transactions let us grant/revoke access, fund/defund vaults, and settle usage in a single tx—auto-pausing keys when balances or limits are hit.